AI Boosts Manufacturing Particle Detection

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.May 16 2024

In an article published in the journal Nature, researchers proposed a novel framework combining deep learning and preprocessing algorithms for particle detection in complex manufacturing backgrounds.

Traditional methods struggled with heterogeneous backgrounds, but this framework adapted to diverse environments, employing preprocessing, artificial intelligence (AI) model selection, and post-processing. Demonstrated with particle entrainment, the system achieved consistent detection accuracy across various conditions, showing promise for improving in-situ process monitoring in manufacturing operations.

Background

Powder particles and granular materials are pivotal in various manufacturing processes, influencing product properties like conductivity, roughness, and porosity. Yet, accurately characterizing particle size distributions amidst complex manufacturing backgrounds remains a challenge. Traditional methods, including size analyzers and image processing techniques, struggle with heterogeneous environments typical of manufacturing settings.

While software tools like ImageJ offer some capability, they often require manual parameter adjustments and lack precision in detecting particles against varied backgrounds. Moreover, existing deep learning approaches primarily excel in homogeneous backgrounds, limiting their effectiveness in manufacturing contexts.

This paper addressed these gaps by proposing a novel framework that integrated deep learning with pre-processing and post-processing algorithms for particle detection in heterogeneous backgrounds. By combining basic image processing with advanced deep learning, this approach offered a semi-supervised solution, reducing the reliance on extensive technical expertise.

The framework aimed to enhance particle identification accuracy in manufacturing scenarios, facilitating in situ process monitoring and advancing particle analysis in complex environments. Through experimentation and comparison with traditional methods, this research demonstrated the efficacy of the proposed framework in addressing the challenges posed by heterogeneous particle–substrate interfaces encountered in manufacturing operations.

Sample Preparation and Image Capturing

The researchers utilized a continuous dipping system to mimic a manufacturing process, entraining particles onto substrates. Initially, a polymer solution was formulated using polymethyl methacrylate (PMMA) and 1,3-Dioxolane solvent. After thorough stirring, a uniform liquid carrier solution (LCS) was obtained. Particles were introduced into the LCS to create a dipping mixture, maintaining a Newtonian regime ratio. Cylindrical rods served as substrates, dipped into the mixture for particle entrainment, preventing sedimentation through agitation.

As the rods were withdrawn, a balance between viscous drag and capillary action facilitated particle transfer onto the substrate. Once entrained, particles adhered using the binder as glue, with rapid solvent evaporation. In-situ analysis was conducted using a microscope or camera, showcasing a heterogeneous morphology. Five types of heterogeneous images were created by varying particle, substrate, and imaging processes, captured at different ambient lighting conditions. Images were taken with a digital microscope, covering a defined area, and used for framework evaluation, training, and validation.

Design and Implementation of the Proposed AI Framework

The framework consisted of four main steps: preprocessing, model selection, AI-guided particle detection, and post-processing. Preprocessing involved enhancing and sharpening images to improve detection accuracy by adjusting contrast and brightness. Model selection determined the appropriate AI model for each particle–substrate image, using transfer learning with MobileNet as a base model. AI-guided particle detection employed algorithms like you-only-look-once (YOLO)v5, chosen for its ability to detect small objects efficiently.

Training multiple AI models for different particle–substrate combinations ensured optimal performance. Post-processing aimed to reduce noise and false positives using domain knowledge. Algorithms pruned excessively small or large bounding boxes, removed overlaps, and eliminated skewed shapes. These steps enhanced overall efficacy by refining detections and improving precision.

The framework's modular design allowed for flexibility and future additions, promising robust particle detection across diverse manufacturing scenarios. Each step contributed to a comprehensive approach, combining advanced AI techniques with practical domain expertise to address complex challenges in particle detection and analysis.

Experimental Setup and Results

The researchers employed a robust experimental setup featuring a high-performance desktop computer equipped with an Intel Core i9 central processing unit (CPU), NVIDIA GeForce RTX 3060 graphics processing unit (GPU), and PyTorch framework for accelerated training. Five diverse particle-substrate image datasets were prepared, annotated, and divided into training, validation, and test sets using Roboflow. Through meticulous preprocessing and augmentation techniques, researchers ensured the readiness of the datasets for YOLOv5-based particle detection.

After training the YOLOv5s model with varying configurations, the authors conducted comprehensive analyses, including ablation studies, to assess the impact of preprocessing and post-processing on detection performance. Notably, the results highlighted the critical role of both stages, with preprocessing enhancing model precision and post-processing refining overall accuracy. Particularly in datasets with higher substrate heterogeneity, these steps proved instrumental in achieving balanced recall and precision.

By integrating transfer learning-based model selection, our framework exhibited consistent performance across datasets, demonstrating its scalability and efficiency. Notably, the proposed approach achieved impressive processing speeds, underscoring its practical viability for real-time applications, with an observed efficiency of 80.86 frames per second during inferencing.

Conclusion

In conclusion, the proposed framework offered a groundbreaking solution for particle detection in complex manufacturing backgrounds. By seamlessly integrating deep learning with preprocessing and post-processing algorithms, it achieved consistent and accurate results across diverse conditions. This innovation promised to revolutionize in-situ process monitoring, enabling real-time adjustments and enhancing quality control in manufacturing operations.

With its adaptability and efficiency, the framework held immense potential for various industrial applications, ranging from three-dimensional printing to semiconductor manufacturing and beyond, ushering in a new era of precision and efficiency in particle analysis and process optimization.